Margin Line Deformation Network Based on Graph Convolution Combined with with Geometric Consistency

Abstract

To address the problem of the high computational complexity of the curved surface deformation method, ignoring the relevance of non-local information on the curve deformation method, and the regardless of the geometric consistency between the position and normal, a graph convolutional layer that simultaneously models local and non-local information is proposed, and a geometry aided prediction module that uses position and normal geometric constraints is designed. Firstly, the <italic>k</italic>-nearest neighbor algorithm and attention mechanism are used to mine local and non-local information to predict the position and normal information of points in space. Secondly, the position of the predicted points is input into the geometry aided prediction module, then the geometric consistency constraint is used to calculate geometric normal information. Finally, the geometric normal information is fused with the normal information predicted by the graph convolution network, then the graph convolution network is reused to obtain the final deformation prediction result. The experimental results of self-built margin line dataset comprising of 876 samples show that this method reduces the mean square error measurement by 16.7% when it compares with other mainstream methods.